Gemini at Work 2026: Google's New Agent Gets Its Own Email and Works With Claude

Thursday, October 8, 8:30 PM Israel time

Hangar One at Moffett Field, California, was built in the 1930s for airships. Now Thomas Kurian, CEO of Google Cloud, is standing in it and presenting what Google calls "your new single, universal agent for work." A few minutes later it turns out the agent gets an email address, a calendar and a Drive folder, and that it also works with Anthropic's models.

Google didn't unveil a new model at the event. It unveiled a project manager that decides on its own which model to run for each task.

Gemini at Work 2026 was the event where Google separated the agent from the model. The new Gemini agent takes a goal, breaks it into steps, brings in subagents and keeps working in the cloud after you've closed your laptop. For each task it picks the right model, from the Gemini family or from Claude's models. We read Google's official announcements, the documentation and the press coverage. Here you'll find exactly what was announced, what's already available and what isn't yet. At the end comes the question that matters: which of this is relevant to a small business.

In short

The Gemini agent is the headline announcement of Gemini at Work 2026: a standing work agent that plans, runs subagents and keeps working for hours and days in the cloud. It works inside Gmail, Docs, Sheets and Drive, and also in Microsoft 365 and Slack.
Every "coworker" agent has its own Workspace account, with email, calendar and Drive. It acts under its own identity and sees only what you've shared with it.
According to Google, the agent already runs each task on the model that fits it, from the Gemini family or from Claude.
Price, the list of plans and an availability date haven't been published. Small businesses currently have early access, and according to Google the agent will reach all customers "soon."

The general contractor: the idea that explains the whole event

Every big renovation has a general contractor. They don't lay the tiles or wire the electricity themselves. They take your plan, decide who does what and in what order, bring in subcontractors and make sure the job gets finished. If the best electrician in town gets replaced, the project goes on. The plans, the keys and the memory of what was agreed stay with the contractor.

That's exactly what Google presented. Kurian put it this way on the official blog: "Gemini is the agent, and the model underneath it is a separate choice". His reasoning: the leading model changes every few months, so when the choice stays open, your context, skills and data don't move.

Our illustration, based on the official Google Cloud blog. The subagent names are examples. Google describes subagents as temporary, each one dedicated to a single job.

This way of thinking isn't new to us. It's how we've built agents from day one, and it's also why we set up Sigma AI Ranking, to know which model fits which job. What's new is a tech giant saying it from the stage, and recommending a competitor's models itself.

What was announced at Gemini at Work 2026: the full table

Announcement	What it is	Availability	Source
Gemini agent	A standing work agent: plans, runs subagents, remembers, and works in the cloud for hours and days	Early access for small businesses. "Soon" for all customers. No price and no date	Google Cloud blog
Coworker agents	An agent in a standing role, with its own Workspace account, email, calendar and Drive	No separate date published	Google Cloud blog
Claude inside the agent	Routes each task to the model that fits it, Gemini or Claude	According to Google, already today	Google Cloud blog
Workspace Intelligence	Spots tasks that can be handed to the agent, and hands them over in one click	Part of the agent	Google Cloud blog
Agent Sandbox and Agent Gateway	An isolated runtime environment, and a firewall that all agent traffic passes through	Status not stated	Google Cloud blog
Spend caps and smart routing	A hard cap on AI spend per project, and choosing the cheapest model that fits the task	Spend caps were already announced in August, and the event added tools on top of them	Google Cloud blog, VentureBeat
Industry editions	Financial services and legal in preview. Government, healthcare and retail "soon"	Preview	Google Cloud blog
AI Solution Finder	A catalog of third-party agents and connectors, including Shopify, Wix, Notion and Docusign	Live	Google Cloud blog

Three things that weren't at the event: a new model, a price and a general availability date. Gemini 4 Argon, Google's most powerful model, was already announced on September 30 and still isn't open to everyone.

What the Gemini agent actually does

According to the official blog, it's a single prompt box that answers questions but mostly does work: research, documents, media, writing code and running it. According to TechCrunch, Kurian described it this way: you give it "objectives, not just instructions".

Works when you don't

A task, a schedule or a response to an event
Work lasting hours and days continues in the cloud
"You never have to re-brief it"

Divides the work

Sets up temporary subagents
Each with its own identity
Steps in parallel and in sequence

Remembers and learns

Four kinds of memory: conversation, knowledge, procedures and events
"Skills" it writes for itself and reuses
Onboards itself "the way a new hire would," according to Google

Connects to tools

Workspace, Microsoft 365, Slack, Salesforce, Jira
Local files on your computer
Any MCP server

What makes it useful is where it happens. The agent works inside Gmail, Drive, Docs, Slides, Sheets, Chat and Calendar, and also on the web, in the mobile and desktop apps, on the command line, in Microsoft 365 and in Slack. According to TechCrunch there's also a "task inbox" that shows how it's thinking, who it handed work to and how far along it is. That part doesn't appear on Google's blog.

A coworker with its own email address

This is the part most different from anything we've seen so far. A coworker agent is a standing agent in a defined role. You describe the role, and Gemini sets it up. According to Google, it gets a full Workspace account: an email address in the style of @agents.company.com, a calendar, Drive and a place in the company directory. It acts under its own identity, not yours, and sees only what you've shared with it. When it edits a document, the change appears under its name in the version history.

For small businesses, Google gave example roles: project manager, bookkeeper, inventory manager and marketing advisor. And these are the four examples it showed of the work itself:

Booking a meeting without giving a single name

Set up a meeting next week with the usual regional event leads.

According to 9to5Google, Gemini finds the attendees on its own, from a Chat space and a thread from a previous event. It checks calendars and emails attendees outside the organization.

✓ Not one name and not one email address

A task that arrives by email

The manager writes: "Send me a project update, as a slide deck."

Workspace Intelligence recognizes this as a task it can hand off, and offers to pass it to Gemini in one click. It also flags the most important message in the inbox, not just the latest one, and explains why.

A coworker agent in a Chat group

A marketing manager asks the event coordination agent for a launch readiness document.

The agent writes the document and posts it to the group when it's ready. In documents it suggests edits and replies to comments, always under its own name.

A task that moves across apps

Research the market trends, build a financial model and prepare a deck.

The agent does the research, builds the model in Sheets and prepares a deck that presents both. It's a task that moves across three apps without you passing a file along by hand.

The examples come from the Google Cloud blog and 9to5Google's coverage. The wording in the bubbles is ours.

What does this mean for a small business owner? The idea of "an employee with its own account" is right, and we've been recommending it for a long time: an agent should have a separate identity and its own permissions, not your password. The open questions are practical. VentureBeat and other sources ask whether every coworker agent consumes a paid Workspace license, and who in the organization is allowed to create one. Google hasn't answered yet.

Claude inside Gemini: why Google runs a competitor's models

This is the line that surprised us most at the event. The official blog says the agent runs each task on the model that fits it, and that as part of this it already orchestrates between the Gemini family and Anthropic's Claude models. According to Google, more models, closed and open, will join later. The stated goal is better quality and lower cost.

The context: on Google's agent platform you can already choose Claude Sonnet 5.5, since September 28, and Claude Haiku 5.5, since October 7. Google isn't pretending there's one model that's best at everything. It's selling the layer above the models: memory, skills, connections and security.

We compare the models themselves in our article ChatGPT, Gemini or Claude.

Security: Kurian's four questions

Kurian built the security segment around four questions. These are the questions every business owner should ask about any agent, from Google or any other vendor:

The four questions, and Google's answers

Who is the agent?

Its own identity, cryptographically attested
Governed "like an employee"

What is it allowed to do?

Role-based, least-privilege permissions
Approval from the organization's security admins

What did it do?

Every action is recorded in an audit log
The action is logged under the agent's name, not a person's

What should it never touch?

Agent Gateway: a firewall for all agent traffic
For example: blocking documents classified Need to Know

Source: Google Cloud blog, October 8, 2026. All agents run inside Agent Sandbox with their own network boundary.

It's a good framework, and it looks like it was written for corporate IT managers. What's missing is the question a small business owner cares about most: what does the agent do without asking you? Before the event, Android Authority reported a toggle called "Require a human review" spotted in a test build. The event blog doesn't mention that toggle. Until Google explains how to require human approval for sensitive actions, like emailing a customer, making a payment or deleting something, that's the first thing to check. We went deeper on this in our article on AI agent permissions and security.

How much does it cost? No answer yet

Google hasn't published a price for the agent and hasn't said whether it will be included in existing Gemini Enterprise or Workspace subscriptions. What it did present were tools for controlling cost: automatically routing each task to the cheapest model that can do it, and a hard spend cap per project in Cloud Billing. When the cap is reached, that project's agent pauses, and you can resume the work with a click. According to VentureBeat, the caps themselves were announced back in August.

98%drop in price per token since 2024, according to Google

5x+growth in small businesses' use of Google Cloud AI tools over the past year

~80%of Google Cloud customers use its AI products

Source: Google Cloud blog and its small business blog, October 8, 2026. These are figures Google reports about itself, with no external verification.

The first number matters more than it seems. When the price per token falls this fast, the question is no longer "how much does a model cost" but how much work the agent does, and in how many attempts. An agent that works for days on end can rack up costs even when every token is cheap. That's why a spend cap is a must-have, not a nice extra. We broke down how agent pricing is built in our article how much an AI agent costs.

The same week: Nano Banana 2.1 and the free tier cut

Two changes happened around the event, and both affect small businesses more than any of the enterprise announcements.

Nano Banana 2.1 has been available to everyone since October 6. It's Google's new image model. According to Google, it delivers better quality, follows instructions more closely and renders text better. It's available in the Gemini app, Google AI Studio, the API, Google Ads and Flow. In the API, the price per token was cut in half compared with Nano Banana 2:

Price per image in the API, standard tier, in dollars

Nano Banana 2.1 · 1K

$0.0336

Nano Banana 2 · 1K

$0.067

Nano Banana 2.1 · 2K

$0.0504

Nano Banana 2 · 2K

$0.101

Nano Banana 2.1 · 4K

$0.113

Nano Banana 2 · 4K

$0.151

Source: Gemini API pricing page, October 9, 2026. Batch processing costs half as much. There's no free tier in the API. Some press reports listed a lower price for 4K, but Google's pricing page states $0.113.

What about the old model? Nano Banana 2 has been deprecated, but according to Google's deprecations page, no shutdown date has been set yet. A date was reported in the press, but that's not what the official page says.

From October 9, free Gemini works only with Flash-Lite. According to Google's help page, the change starts today for users without a subscription:

Plan	Models in the Gemini app
No subscription	Flash-Lite only
Google AI Plus	Flash-Lite and Flash, with double the quota
Google AI Pro	Flash-Lite, Flash and Pro, with 4x the quota
Google AI Ultra	Flash-Lite, Flash and Pro, with the highest quota

According to 9to5Google, this is Gemini 3.5 Flash-Lite. According to reports, the change applies only to personal accounts, not to Workspace and not to the API. The timing isn't uniform either: AI Plus subscribers will get an email with their own date, and Google's subscriptions page still shows the previous free tier. Google is also adding a choice of effort level: low, medium or high. A higher level uses more of the quota.

What does this mean for you? If people in your business use Gemini on free personal accounts for writing, summaries or research, you may notice a drop in quality as soon as this week. It's a good time to check who on the team uses what, and whether a properly set up business account would be better.

Three companies, one idea, within two weeks

Google isn't alone. Over the past two weeks, each of the big companies has introduced a standing work agent. And all of them are still in early access, preview or beta.

Two weeks of announcements. Sources: the official Microsoft, Google and Anthropic blogs, and our DevDay coverage.

Product	Status	Where you use it	What sets it apart
Gemini agent, Google	Early access, no price	Workspace, Microsoft 365, Slack, web and apps	A Workspace account for every coworker agent, and routing between Gemini and Claude
dots, OpenAI	Rolling out gradually to Pro, Business Premium and Enterprise in select markets	ChatGPT, Slack, Microsoft Teams	A cloud computer and browser, and rules for what's allowed, what needs approval and what's blocked
Autopilot, Microsoft	Private preview	Copilot	According to Microsoft, its own "identity, memory, computer and workspace." Agent work is billed by usage
Claude for Google Workspace, Anthropic	Public beta for paid subscribers	Docs, Sheets, Slides	A mode where Claude asks before it edits

The direction is clear: everyone is selling an employee who never goes home, and everyone is trying to sit inside the tools you already work in. Google's advantage is that it's already there. If you run your business in Gmail and Drive, you get the agent right where the work already is. We summarized OpenAI's announcements in our DevDay 2026 article.

Should you wait for the Gemini agent?

Five questions, and you'll get an answer.

What about your language and your country? Google hasn't stated which countries or languages the new agent will support. If you're in Israel, here's what is known: according to the Workspace language support page, the Gemini side panel in Gmail, Docs, Sheets and Drive already supports Hebrew, while "Help me write" in Gmail and Docs isn't on the list yet. Wherever you are, check availability in your account and test how the new agent handles your language before you rely on it.

Hype versus reality: what we think

What you heard from the stage

The headlines

"Your new single, universal agent for work"
Works for hours and days without you re-briefing it
Any model, any tool, any MCP server
Available to all customers "soon"

What's worth checking

The bottom line

Which actions it takes without asking you
How often a long task really finishes the way it should
Which data terms apply when the task goes to Claude
Price, plans and date, which haven't been published

Our opinion: the important announcement at the event isn't the agent itself, but the admission that the agent and the model are two separate things. When Google, which invests billions in its own models, writes that the leading model changes every few months, that tells you something simple. Don't choose an AI vendor based on whichever model is strongest this month. Choose based on who holds your context, your connections and your approval rules, and knows how to switch models when needed.

What Google still lacks is data. There's no independent test of how well the agent completes long tasks that move across apps, and VentureBeat raises that question too. The customer numbers, like "5x" among small businesses, are figures Google reports about itself. And according to VentureBeat, the customer examples shown at the event don't necessarily use the new agent.

What this means for Platin Digital clients

Platin Digital is an OpenAI Select Partner, and the agents we set up are built with whichever models and tools fit each business best, from OpenAI, Anthropic, Google and others. The approach Google presented, one agent on top of several models, has been our approach from the start.

Google's models get tested like any other model. When access to Gemini 4 Argon opens in the API, we'll test it on real tasks, and only then decide whether it goes into the work, and where.
The approval rules don't change. Publishing, budget, deletion and sending a message on behalf of the business continue to wait for your approval, on every model and with every vendor.
If the Gemini agent fits you, we'll say so. A business whose work is all in Workspace may get a lot out of it. A business whose work is on a WordPress site, in Google Ads and in Meta needs a direct connection to those systems. Want to understand the principle first? Start with what an AI agent is.

Frequently asked questions

What is Gemini at Work 2026?

Gemini at Work 2026 is a Google Cloud event held on October 8, 2026 at Hangar One at Moffett Field, California. The keynote was given by Thomas Kurian, CEO of Google Cloud. The headline announcement was the Gemini agent, a standing work agent for organizations and businesses, alongside coworker agents with their own Workspace accounts, security tools for agents, cost control tools and industry editions.

What is the difference between the Gemini agent and the regular Gemini app?

The Gemini app answers you in a conversation. The Gemini agent takes a goal and carries it out: it plans steps, runs subagents, uses tools and skills, and keeps working in the cloud for hours or days even when your computer is off. It also picks the right model for each task, including Claude models.

When will the Gemini agent be available, and how much will it cost?

Google hasn't published a price or a date. According to the Google Cloud small business blog, the agent is in early access with small business customers and will be available to all customers "soon." According to 9to5Google, it will come to select Workspace Business and Enterprise plans, but the list of plans hasn't been published.

Does the Gemini agent work in my language and my country?

Google hasn't stated which countries or languages the new agent supports. In Israel, the Gemini side panel in Gmail, Docs, Sheets and Drive already supports Hebrew, according to the Workspace language support page. For the agent itself, wait for an official announcement, or check availability in your account when it opens and test how it handles your language.

Why does Google use Claude models inside Gemini?

According to the Google Cloud blog, the agent runs each task on the model that fits it, to improve quality and lower cost, and it already chooses between the Gemini family and Anthropic's Claude models. Google explains that the leading model changes every few months, so the choice of model stays open, and the customer's context and data don't move.

What changed in Gemini's free tier on October 9?

According to Google's help page, from October 9, 2026, users without a subscription get only the Flash-Lite model in the Gemini app. Google AI Plus subscribers get Flash-Lite and Flash, and AI Pro and Ultra subscribers also get Pro. According to reports, the change applies to personal accounts, not to Workspace and not to the API.